How to Connect Evidence Across Literature Review Sections

A literature review rarely fails because the sections are empty. It fails because the sections do not talk to each other: the background cites one set of papers, the methods section cites another, the discussion introduces claims that never appeared earlier, and the final recommendation rests on a study that was screened out three drafts ago. Connecting evidence across sections means building a traceable line from each claim back to a specific source, and forward from each source to every section that relies on it.

This guide gives you a practical sequence for that work, aimed at teams producing client decks, decision memos, team handoffs, or project closeouts. It covers how to diagnose where your review currently breaks, which artifacts to build, how to run the connection pass section by section, and how to know the result is ready to ship.

Before you reorganize anything, find out what kind of disconnection you have. Most reviews fail in one of four ways, and each needs a different fix.

• Orphan claims. A sentence in the discussion or recommendation asserts something no earlier section established. The fix is to trace the claim back to a source or remove it.
• Orphan sources. A paper sits in the reference list but no section actually uses it. Either it belongs somewhere and was forgotten, or it should be cut.
• Section silos. Each section was drafted by a different person with a different search, so the background and the methods draw on non-overlapping literatures. The fix is a shared evidence base, not more editing.
• Provenance gaps. The claim is supported, but nobody can say which study, which report of that study, or which version of the source it came from.

A quick diagnostic: pick five load-bearing claims from your conclusions and try to trace each one backward through the sections to a named source. If you cannot complete the trace for most of them, you have a provenance problem, and no amount of prose polishing will fix it.

Connection is not a single standard. Decide up front how strict your trace needs to be, because that decision determines how much structure you build.

Write this standard down before you start. It becomes the acceptance test at the end.

You cannot connect sections that were built on different searches. The fix is a single evidence base that every section draws from.

Cochrane's guidance on searching is a useful model even outside systematic reviews: plan searches from your eligibility criteria, use multiple sources, combine free-text terms with subject headings, document the process, deduplicate records, screen in stages, and link multiple reports of one study together. Each of those steps produces an artifact you will reuse when connecting sections.

In practice, build four artifacts:

1. A search log. Which databases and interfaces you searched, the exact terms and filters, the date run, and the number of records returned. PubMed's search details view, which shows how terms were translated and combined, is a good example of the level of detail to capture. Google Scholar's date limits and sorting options are the kind of setting worth recording so a rerun is reproducible.
2. A screening log. For each record: included, excluded, or pending, with a one-line reason. Cochrane's staged screening approach maps directly onto this.
3. A source register. One row per source, with a stable identifier, the study it belongs to, and any linked reports. Cochrane's point about linking multiple reports of one study matters here: a trial and its follow-up paper are one evidence item, not two.
4. A claim register. One row per claim you intend to make, with the source or sources behind it, the sections where it appears, and a confidence level.

The claim register is the artifact that actually connects sections. Everything else supports it.

If your team needs a place to save and organize the papers behind these registers, My Library lets you save or upload papers and keep them organized as a working collection. That collection becomes the backing store for your source register.

With the registers in place, work through the review in a fixed order. Do not edit prose yet; you are mapping, not rewriting.

Read each section and highlight every sentence that asserts something about the evidence. Assign each a claim ID. Most reviews have far fewer load-bearing claims than sentences; a 4,000-word review might have 30 to 50.

For each claim ID, record the source or sources. Where a claim rests on a single study, say so. Where it rests on several, list them. Where you cannot find a source, mark it as unsupported and flag it for the next pass.

Invert the register: for each source, list every section that cites it. This is where orphan sources and section silos become visible. A source cited only in the discussion but never in the background is a signal that the background is incomplete.

For each recommendation or conclusion, walk backward: conclusion, to claim, to source, to the screening decision that admitted the source. For each source, walk forward: source, to the claims it supports, to the sections where those claims appear. Both directions should complete without gaps.

Gaps come in three kinds, and each has a defined action:

• Claim with no source. Either find a source through a targeted search or downgrade the claim to an observation about the reviewed literature rather than a finding.
• Source with no claim. Either integrate it where it belongs or remove it from the reference list.
• Claim supported only by an excluded or pending source. Re-screen the source or remove the claim.

Cochrane recommends rerunning searches before publication. For a client deck or decision memo, the equivalent is a final search pass close to the delivery date, logged the same way as the first. If new records appear, screen them and update the registers before the connection pass is considered closed.

By the time you reach the connection pass, you are not doing a fresh literature search. You are filling named gaps. That changes how you search.

• For a missing source behind a specific claim, search by the claim's key terms and by the authors or title fragments you already know. Google Scholar's author and title search, and PubMed's field tags and citation lookup, are built for exactly this kind of targeted retrieval.
• For a source you suspect exists but cannot locate, use citation chasing. Google Scholar's Cited by and Related articles, and PubMed's Similar Articles, let you move from a known paper to its neighbors.
• For a claim that may be outdated, sort by date or apply a date limit, and record the limit in your search log so the constraint is visible in the final deliverable.

When you need to run these targeted searches inside a single scholarly interface, WisPaper search is the entry point for finding academic papers. For a public explanation of the same capability, the Scholar Search page describes finding academic papers.

Suppose a team is preparing a decision memo recommending whether to expand a screening program. The draft has four sections: background, evidence on outcomes, implementation considerations, and recommendation.

The problem. The recommendation cites two studies. One appears nowhere else in the memo. The background cites three studies that never reappear. The implementation section makes a claim about cost that has no source at all.

The registers. The team builds a source register with five sources and a claim register with twelve claims. Mapping sources to sections shows that two of the background sources support the same claim as one of the recommendation studies, which means the recommendation is better supported than the draft suggested. The cost claim has no source and is downgraded to an open question for the implementation team.

Confidence levels. Each recommendation gets an explicit label. The outcome claim, supported by two linked reports of the same study, is marked high confidence with a note that both reports come from one study. The implementation claim is marked medium, because it rests on a single source. The cost question is marked unresolved.

The trace. The final memo includes an appendix mapping each recommendation to its claims and sources, with the screening decision for each source recorded. A reader can walk from the recommendation to the source in two steps.

The closeout. The team reruns the search two weeks before delivery, finds no new eligible records, logs the rerun, and closes the connection pass.

This example shows the core principle: connection work is mostly bookkeeping, and the bookkeeping is what makes the memo defensible.

A connected review can still break at the reference list. Hallucinated or mismatched references are a common failure mode when drafts are assembled quickly or with automated assistance.

Before the final connection pass, verify that every entry in the reference list corresponds to a real paper. TrueCite checks BibTeX against real academic databases to flag hallucinated references, which fits the moment when you have a complete reference list and need to confirm it before mapping claims to sources. Run this check before the claim-to-source mapping, not after, so you are not tracing claims to references that do not exist.

Once your sources are collected and organized, some connection questions are best answered by querying the collection directly: which of my saved papers address this claim, what does this set of papers say about a specific outcome, or where do two sources disagree.

Library QA supports asking questions based on papers uploaded to My Library or added from search results. Use it during the gap-resolution step, when you have a specific claim and need to know whether your existing collection already supports it. Record the answer and the papers it draws on in the claim register, exactly as you would for a manual read.

The connection pass is not a one-time cleanup. It has to survive the handoff.

• Handoff artifact. Deliver the claim register alongside the review. A new owner should be able to see which claims are load-bearing, which sources back them, and which are flagged as unresolved.
• Change log. When a claim changes, record which source changed and which sections were updated. This prevents the classic failure where a source is swapped in one section and the old claim survives in another.
• Closeout checklist. Before closing the project, confirm that every source in the register is used, every claim has a source or an explicit unresolved flag, and the final search rerun is logged.

If your team runs multiple reviews or wants the connection work to persist across deliverables, Research Projects is the area for creating and managing a project. Keeping the registers inside a project means the next review starts from an existing evidence base rather than a blank page.

Use this checklist as your acceptance test. The review is ready when every line is true.

• Every load-bearing claim has a claim ID and at least one named source, or an explicit unresolved flag.
• Every source in the reference list is cited in at least one section, or has been removed.
• Multiple reports of the same study are linked as one evidence item in the source register.
• Each recommendation traces backward to a source and forward to the sections that use it.
• Confidence levels are assigned to every recommendation and every claim that drives one.
• The search log records terms, filters, dates, and the final rerun.
• The screening log records include/exclude decisions with reasons.
• The reference list has been checked for hallucinated or mismatched entries.
• The claim register is delivered with the review as a handoff artifact.

If any line is false, you have a named gap and a defined next action. That is a better state than a polished draft with invisible breaks.

Connection work can run indefinitely if you let it. Set explicit stopping rules.

Stop when:

• Every load-bearing claim traces to a source or carries an unresolved flag.
• The final search rerun returns no new eligible records, or new records have been screened and logged.
• The claim register and source register are complete enough for a new owner to use without asking you questions.

Escalate when:

• A recommendation depends on a single source and no corroborating source can be found. Escalate to the decision owner rather than presenting it as settled.
• Two sources conflict and the conflict cannot be resolved from the available evidence. Record both and escalate the conflict rather than picking one silently.
• A claim cannot be sourced and cannot be downgraded without changing the recommendation. That is a decision for the owner, not the reviewer.
• The search rerun surfaces a study that would change the recommendation. Pause delivery and rerun the connection pass.

These rules protect the review from two opposite failures: shipping with invisible gaps, and never shipping because the evidence base could always be larger.